Design Of Special-Purpose Arrays For Matrix Computations: Preliminary Results.
Jaime H. Moreno, Tomás Lang · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1987
We describe some preliminary results in the development of a general and systematic methodology to design arrays of processing elements (PEs) for matrix computations, with the capability to handle algorithm and implementation in a unified manner. This is a transformational methodology, based on the dependence graph of the algorithms. It provides mechanisms to deal with issues such as data broadcasting, data synchronization, interconnection structure, I/O bandwidth, number of PEs, throughput, delay, and utilization of PEs. We show that different transformations may lead to entirely different computing structures and that the selection of suitable transformations is directed by the specific restrictions imposed on the implementation. We apply a preliminary version of this methodology to the algorithms for matrix multiplication and LU-decomposition. The approach produces structures which correspond to proposed systolic arrays for these computations, as well as structures that exhibit better efficiency than those arrays.